I wrote a long blot post on the economics of AI, prompted by two great workshops (Windfall and NBER), and me leaving OpenAI for METR:
TLDR: we driving in the fog.
1. There is no standard model of AI’s economic impact. Economists have been using a wide range of assumptions to
I made a list of forecasts of the impact of AI on economic growth over the next decade.
A few observations... (🧵):
tecunningham.github.io/posts/2025-10-…
Stylized facts about AI capabilities that economists should know:
1. We have no standard metric for computer intelligence. Many metrics have been proposed but none have been adopted as a standard. This makes talking about capabilities difficult, it is intellectual quicksand.
I wrote a long evaluation of Meta’s 2020 election experiments: the small effects on affective polarization found in these experiments are consistent with large effects of social media on society at large. I.e., I’m not sure how much we’ve learned.
2. GDP will be a poor proxy for AI’s impact. AI’s benefits are likely to elude GDP for two reasons: (1) it will reduce the necessity for exchange (and GDP measures exchange); (2) it will lower the labor required for services, and the value-added from services are typically
A new note on ranking by engagement tecunningham.github.io/2023-04-28-ran… - six main observations:
(1) Most platforms (FB, IG, Twitter, TikTok, YouTube, Netflix) rank content by predicted engagement: they'll show you the items most likely to cause you to click, reply, retweet, etc.
My paper with Josh Kim on interpreting multi-outcome experiments -- git.io/JegrC -- (1) multivariate shrinkage, (2) good news is bad news, (3) surrogate metrics, (4) causal effects, (5) composite metrics. / #codecon19
A long and partisan note about experiment interpretation based on experience at Meta and Twitter.
The common thread is that people are pretty good intuitive Bayesian reasoners, and so the best thing to do is to summarize the relevant evidence and let a human be the judge. (🧵)
4. AI will likely have a discontinuous impact on science and technology. Many existing models treat computers as substitutes for humans in the R&D process, but there is reason to expect AI to have a qualitatively different effect on scientific progress, than just substituting
3. Transformative AI will raise the relative value of resources, and possibly lower the value of labor. If computers can do all human work then there will still be scarcity in natural resources (land, energy, minerals). Because humans require resources to do work (energy, land),
6. Inference prices are falling at 10X/year. The price of an LLM of a given level of capability is falling at around 10X/year. This is mostly due to better training, not lowered price of compute.
